摘要:Abstract Standard approaches to the diagnostic decision making require prior knowledge about the diagnosed process and data records at various faulty conditions. This is not easy to meet in practice. In this paper we propose a data-driven non-parametric approach to fault detection and isolation, which relies on evaluating the dissimilarity statistic feature evaluated from current measurement records and features taken during the nominal fault-free machine condition. For that purpose the Jensen-Rényi divergence of the empirical distributions is employed. Hence the need for data records at different failure modes is avoided. The approach has been initially developed for robust bearing diagnostics and is presented here in that context. It has been validated on simulated and experimental case studies with bearings faults. Moreover, the approach is rather general and applicable in a number of other domains as well.